Skip to content

Fun-ASR-Nano

Fun-ASR-Nano-2512 is FunAudioLLM's compact ASR model with a SenseVoice-style audio encoder, Qwen3-0.6B text decoder, hotword prompting, and language-conditioned transcription.

Convert

The upstream release ships PyTorch weights in model.pt, so convert it with the model-specific converter:

python -m mlx_audio.stt.models.fun_asr_nano.convert \
  --hf-path FunAudioLLM/Fun-ASR-Nano-2512 \
  --mlx-path Fun-ASR-Nano-2512-mlx

The converted directory is directly uploadable to Hugging Face as mlx-community/Fun-ASR-Nano-2512.

Usage

from mlx_audio.stt.utils import load_model

model = load_model("mlx-community/Fun-ASR-Nano-2512")
result = model.generate("audio.wav", language="zh", hotwords=["开放时间"])
print(result.text)

mlx_audio.stt.generate \
  --model mlx-community/Fun-ASR-Nano-2512 \
  --audio audio.wav \
  --output-path transcript \
  --language zh

Options

language accepts ISO hints for the current Nano checkpoint: zh, Chinese dialect codes that map to the Chinese prompt (yue, wuu, nan, hak, gan, hsn, cjy), en, and ja. Pass None or auto to omit the language constraint.

hotwords accepts a list of domain terms to include in the upstream prompt.

itn=False asks the model to skip inverse text normalization.

Boundaries

The MLX implementation is transcription-only. The public FunAudioLLM/Fun-ASR-Nano-2512 checkpoint does not include timestamp head weights, so this model does not emit word timestamps.

VAD and diarization remain external, so split or diarize audio before calling this model when those features are required.